Disclosed herein are embodiments of an asynchronous memory device that use internal delay elements to enable memory access pipelining. In one embodiment, the delay elements are responsive to an input load control signal, and are calibrated with reference to periodically received timing pulses. Different numbers of the delay elements are configured to produce different asynchronous delays and to strobe sequential pipeline elements of the memory device.
|
11. An asynchronous dynamic random access memory (DRAM) device comprising:
a register to store a delay-select value;
a memory core to output read data in a memory read operation; and
output control circuitry to receive the read data from the memory core and to enable the read data to be output from the asynchronous memory device a predetermined time after receipt of an input load signal, the predetermined time including a delay interval having a duration selected by the delay-select value.
1. A method of operation within an asynchronous dynamic random access memory (DRAM) device, the method comprising:
receiving a command, an input load signal and a delay-select value;
storing the delay-select value within a register;
retrieving read data from a memory core in response to the command; and
asserting a register-load signal a predetermined time after receipt of the input load signal to enable the read data to be output from the asynchronous memory device, wherein the predetermined time includes a delay interval having a duration selected by the delay-select value.
18. A memory module comprising:
a substrate; and
a plurality of asynchronous dynamic random access memory (DRAM) devices disposed on the substrate, wherein each of the asynchronous DRAM devices comprises:
a register to store a delay control value;
a memory core to output read data in a memory read operation;
output control circuitry to receive the read data from the memory core and to enable the read data to be output from the asynchronous memory device a predetermined time after receipt of an input-load signal, the predetermined time including a delay interval having a duration selected by the delay control value.
2. The method of
3. The method of
4. The method of
5. The method of
6. The method of
7. The method of
8. The method of
determining that the command specifies a memory read operation; and
initiating retrieval of the read data from the memory core in response to determining that the command specifies a memory read operation and at a time controlled by the input-load signal.
9. The method of
10. The method of
12. The asynchronous memory device of
13. The asynchronous memory device of
14. The asynchronous memory device of
15. The asynchronous memory device of
16. The asynchronous memory device of
17. The asynchronous memory device of
|
This is a continuation of U.S. patent application Ser. No. 11/165,797 filed Jun. 24, 2005, now U.S. Pat. No. 7,362,626 entitled “Asynchronous High-Bandwidth Memory Component Using Calibrated Timing Elements,” which is a continuation of U.S. patent application Ser. No. 10/846,220 filed May 14, 2004, now U.S. Pat. No. 6,934,201, which is a continuation of U.S. patent application Ser. No. 10/271,936 filed Oct. 15, 2002, now U.S. Pat. No. 6,788,594, which is a division of U.S. patent application Ser. No. 09/796,924 filed Feb. 28, 2001, now U.S. Pat. No. 6,788,593, all of which are hereby incorporated by reference in their entirety.
This invention relates to high-speed memory systems and devices, and in particular to high-speed memory devices that accommodate pipelined memory access operations.
The memory device shown in
The CAS signal also loads a command or instruction (write or read) into a command register 16. A command decode block 17 interprets the current memory instruction and enables an appropriate driver 18 or 19, depending on whether the memory operation is a write operation or a read operation.
This access is asynchronous since read data appears on the DATA bus after a time that is determined by the DRAM and not by timing signals supplied externally (other than the initial CAS edge that loads the address). The advantage of this approach is simplicity—it is relatively easy to use this memory device. The disadvantage is performance—the number of read operations per unit of time is relatively limited since accessing the memory array and transporting the resulting data on the DATA bus must be done sequentially before the next access can begin.
Instead of RAS and CAS signals, this device uses a single CLK signal, in conjunction with the OP bus, to load row and column addresses into registers 22 and 23. The command register 24 is loaded by the CLK signal as well.
Another difference from the circuit of
The inclusion of two or more independent banks of memory arrays permits more that one memory access to take place at a time. In other words, a second memory access operation can be initiated even before obtaining results of an earlier operation. Registers 25 and 26, in the path of the read and write data, are necessary for this type of overlapped operation. Such overlapped operation is typically referred to as “pipelined” operation or “pipelined” memory access.
This operation is synchronous, in that data output is timed and enabled relative to an externally supplied clock signal. The row and column address registers 22 and 23 form a first pipeline stage, in which addresses are obtained for accessing memory. The read data register 25 forms a second pipeline stage, which is capable of holding memory results even as another memory access operation is initiated in the first pipeline stage. As a result of this technique, the two steps of memory access and data transport are done sequentially in the two pipeline stages of the DRAM. A second memory access could be started after the second CLK edge, overlapping the two operations.
There are two benefits to this technique. First, it permits sequential transactions to be overlapped, increasing the number of read transactions per unit of time. Second, it resynchronizes the transport of the read data—the signals that enable and disable the drivers are timed by the subsequent CLK edges.
As the signaling bandwidth of memory buses is increased, more pipeline stages can be added to the DRAM so that individual data slots are very small. Modern memory designs utilize a high degree of pipelining to support very high transfer rates.
Although pipelining has been essential to achieving high memory access rates, the technology does have disadvantages. High latency is one disadvantage, resulting from the need to quantize internal delays to the externally-supplied clock period. A disproportionally high power requirement is another disadvantage. Power is a concern because a free-running clock dissipates power even when no useful work is being done. Some devices utilize low-power modes in which the clock is gated off, but this creates further latency problems. Furthermore, the power needed while restarting the clock threatens to erase whatever savings might have otherwise been gained by disabling the clock.
The described system includes a master device 31, such as a memory controller, and a plurality of slave devices 32-33, which might comprise memory devices. The master device 31 initiates and controls data exchanges over bus 36. It is located at one end of the bus, referred to herein as the master end of the bus. Slave devices 32-33 are spaced along the remaining portions of the bus, toward its other end.
The bus can be configured in a variety of different ways. For example, bus 36 might include a shared address bus that is used for both row and column addresses. Alternatively, bus 36 might include individual buses, dedicated respectively to row and column addresses. Bus 36 also includes a data bus, which might be dedicated to only data or might be shared between data and address information. Furthermore, the data bus might be uni-directional or bi-directional. Bus 36 further includes a command bus, which again might be dedicated or shared.
The bus includes one or more input load signal lines 37 that carry input load signals. An input load signal is issued by master device 31 and received by slave devices 32-33 to initiate data access cycles such as memory read and write cycles in slave devices 32-33. As will be described in more detail below, the slave devices are responsive to the input load signal to load data at predetermined, asynchronous times following the input load signal. In the case of a read cycle, the slave devices load data from internal storage and present or enable such data on bus 36. In the case of a write cycle, the slave devices load data from bus 36.
This example is a DRAM, but the concepts described herein are applicable to various different kinds of volatile, non-volatile, random access, and read-only memory, including SRAM (static random access memory); flash memory; mask-programmable memory; field-programmable memory; electrically-erasable, programmable, memory; ferro-electric memory; magneto-resistive memory, etc.
Furthermore, while certain aspects of the described circuits utilize asynchronously generated signals, it is contemplated that the described asynchronous techniques might be employed in circuits that also utilize or receive periodic clock signals for certain purposes.
DRAM 50 comprises a plurality of memory arrays or banks 52, each having a plurality of memory cells, which will be referred to collectively as the memory core. This type of memory is addressable by bank, column, and row. Typically, the bank address is incorporated in the row address as the highest several bits. The banks are capable of being independently accessed.
Memory such as this is typically accessed by providing a row address, sensing all the columns of the specified row, and then accessing one or more memory columns of the sensed memory row. Column data is available only after a minimum sense time, measured from the previous sensing operation.
DRAM 50 has one or more address registers 54 and 55 that correspond to row and column addresses, respectively. An input load signal LD, also referred to as an address load signal, is received from an external source such as a memory controller, and is used to load the row and column address registers 54 and 55 from a common address bus ADDR. In addition DRAM 50 has one or more command registers 56 that load received command information from a command bus OP at a time indicated by the received LD signal. Command decoding logic 58 responds to the value loaded in command register 56 to gate signals appropriately within the device.
Data is received from a data bus, labeled DATA in
Appropriate buffers 60 are provided for incoming signals.
The memory device includes a read data register 62 that is positioned to receive and latch data from core memory 52 during a memory read cycle. The output of read data register 62 passes through a read output driver 63 on its way to the DATA bus.
The memory device also includes a write data register 64 that receives data from the DATA bus and provides it to core memory 52. A write driver 65 is positioned between write data register 64 and the core memory 52.
Read data register 62 loads memory data from memory cells of core memory 52 at a time indicated by a data register load and enable signal LOAD/ENABLE. More specifically, both read data register 62 and read driver 63 are enabled in common by the LOAD/ENABLE signal. In response to this signal, read data register 62 latches any data being provided by core memory 52, and read driver 63 turns its outputs on to present read data on the DATA bus.
Write data register 64 similarly loads memory data at a time indicated by its received data register load and enable signal LOAD/ENABLE. Specifically, both write data register 64 and write driver 65 are enabled in common by the corresponding LOAD/ENABLE signal. In response to this signal, write data register 64 latches any data being provided from the DATA bus, and write driver 65 turns its outputs on. During a subsequent, independent operation, the data provided from data register 64 is loaded into memory cells of core memory 52.
The LOAD/ENABLE timing signals are created by respective asynchronous delay elements 70 and 71. Each of these elements asynchronously generates its LOAD/ENABLE signal at a predetermined time after receiving the load signal LD. More specifically, command decoding logic 58 is configured so that delay element 70, which is associated with a read operation, receives the LD signal when it is received in conjunction with a column read command from the OP command bus. Delay element 70 responds by delaying the LD signal to create a LOAD/ENABLE signal which is supplied to read data register 62. Delay element 71, which is associated with a write operation, receives the LD signal when it is received with a column write command from the OP command bus. It responds by delaying the LD signal to create a LOAD/ENABLE signal which is supplied to write data register 64.
Delay element 70 is responsive to its input signal to delay its input signal by a predetermined amount tCAC. This time corresponds approximately to the time required from specifying a column address to the time at which the corresponding data is available from core memory 52. Delay element 71 is responsive to its input signal to delay its input signal by a predetermined amount tCWR. This time corresponds approximately to the time required from specifying a column address to the time at which the corresponding write data loaded into register 64 and presented to the memory core 52.
At the rising edge of the LD signal, a received column address is loaded from address bus ADDR into column address register 55 and a read command is loaded from command bus OP into command register 56. The LD signal is passed through command decoding logic 58 and initiates a timing interval within delay element 70. After a time tCAC, the delay element produces the LOAD/ENABLE signal, which both loads memory data from core memory 52 into read data register 62, and also enables output driver 63 (after a delay tON caused by the latency of driver 63). The LOAD/ENABLE signal remains active for a time tBIT, and then turns off driver 63 (after a delay tOFF, again caused by the latency of driver 63). Write cycles occur with similar timing, except that data is latched from the DATA bus rather than from core memory 52.
This configuration allows the two memory operation steps, access and transport, to be performed sequentially, in a pipelined fashion. In a first stage, address and command data are loaded into first stage registers 55 and 56, and memory access is initiated. In a second stage, accessed memory data is loaded into second stage output register 62 and driven on the DATA bus for transport. A second access cycle can be initiated during the transport stage by reasserting the input load signal LD prior to completion of the first access cycle—after a time tCAC from the initial LD signal.
Using this architecture, pipelining is controlled with asynchronous delay elements rather than with synchronous clocking. There are two benefits to this approach. First, the delay of each pipeline stage can be adjusted differently, rather than forcing all of the delays to match an arbitrary clock period. This also reduces latency, especially in cases where the memory controller might operate at a clock cycle time that is not well matched to the pipeline stage delays of the memory.
Power reduction is a second benefit. A prior art synchronous DRAM uses registers to perform two functions: delay and information storage. Furthermore, as a single transaction passes through a synchronous DRAM, all registers must be continuously clocked. The memory device of
Furthermore, as the signaling bandwidth of memory buses is increased, the clock recovery logic (delay-locked loops and phase-locked loops) that are needed for synchronous DRAMs can require circuits that are continuously on and that require a long time interval to reach their operating point. This can result in a significant power level even when no memory accesses are being performed. Turning off these circuits can result in significant added latency when an access must be started.
At the rising edge of the LD signal, column address is loaded from address bus ADDR into column address register 55 and a read command is loaded from command bus OP into command register 56. The LD signal is passed through command decoding logic 58 and initiates a timing interval within delay element 82. After a time t1, the delay element produces the LOAD signal, which loads address pipeline register 81 with the column address from address register 55. After another delay, t2, produced by delay element 70, the LOAD/ENABLE signal becomes active, which loads memory data from core memory 52 into read data register 62 and enables output driver 63. Note that the LOAD/ENABLE signal in this embodiment may be created either by delaying LOAD by t1, or by delaying LD by t1+t2.
Note that the LD signal, which loads addresses and initiates memory access cycles, might take forms other than the simple single-conductor voltage transition shown in the disclosed embodiments. For example, the LD signal might be derived from a combination of two or more other signals that have been logically gated to decode memory access events.
The advantage of this embodiment is the presence of the additional pipeline stage, which allows a higher degree of overlapped operations. In this embodiment, a subsequent memory operation can be initiated sooner than in the embodiment of FIG. 6—at time t1, when the LOAD signal transitions. If desired, additional address pipeline stages can be utilized to provide even higher bus utilization.
This embodiment, and the more complex embodiments that follow, demonstrate the general concept of a memory device having a plurality or sequence of pipeline registers or elements that are asynchronously signaled and loaded, in a predetermined sequence, to complete memory access cycles. In the embodiment of
In the prior art, pipeline elements are generally signaled or strobed by a periodic clock signal, with the disadvantages that have already been noted. In the embodiments described herein, however, a memory cycle is initiated with a single input load signal. Timing logic, including a plurality of delay elements, is responsive to the input load signal to produce a corresponding sequence of asynchronously timed register load signals. This sequence of load signals is utilized to load the various pipeline elements in the proper sequence, with the proper timing.
Delay Elements
The various load signals to the pipeline elements are produced by corresponding delay elements. In the described embodiments, a plurality of delay elements are designed within a single integrated circuit to have matching delays, and individual delay elements are grouped or aggregated to produce delays of different lengths. For example, individual delay elements might be designed to have a delay td. Multiples of td are then obtained by chaining a plurality of individual delay elements.
For high-speed operations, it is desirable to calibrate the individual delay elements as precisely as possible. Such calibration allows external devices, such as memory controllers, to communicate in synchronization with the advance of data through the pipeline elements of the memory device.
The time td is preferably changeable in response to a delay value that is either generated internally to the memory device or received from a source external to the memory device. More specifically, each delay element 90 is responsive to a delay adjustment signal 93. This common signal is provided to all of the matching delay elements. As a result, the delay elements are collectively adjustable, to produce individual matching delays. Such delays are asynchronous—in the described embodiments, the delays are not necessarily aligned to any received clock signal.
The memory device includes delay setting logic 94 that sets and calibrates the delays of the delay elements. Delay setting logic 94 receives a signal 95 that indicates a delay value. In response to the supplied delay value, delay setting logic 94 sets its delay adjustment output 93 to an appropriate value or level, so that each of the delay elements 90 provides the desired signal delay between its input and its output.
Delay setting logic 94 preferably uses a feedback loop to calibrate its delay adjustment output, and to in turn calibrate the delay elements 90. Specifically, one of the delay elements 90a is dedicated for use as a calibration element. Delay setting logic 94 generates a signal at the input of calibration delay element 90a, and monitors the resulting output from element 90a to determine the actual delay resulting from the delay adjustment value 93. Based on this determination of the actual delay, delay setting logic 94 varies its delay adjustment output signal 93 until the desired delay value td is obtained through element 90a. Because the delay elements are all similar in design and implementation, calibrating one of the delay elements ensures that all of them are similarly calibrated.
The delay value 95 can be generated by an internal source such as a capacitive circuit or other type of circuit that is capable of generating a precise reference interval. More desirably, the delay value is generated externally to the memory device, so that the internal delays of the memory device can be synchronized with operations of an external device such as a memory controller.
The delay value 95 can be supplied to delay setting logic 94 in various forms, but is preferably supplied as a pair of signals or timing events that are separated in time by the actual desired delay. For example, the delay value can be specified as the time between two voltage edges on a single input conductor, or as the time between two signal edges on a pair of conductors. Alternatively, the delay value might be specified as two relatively timed events on an input conductor that is normally used for some other purpose, such as a conductor that is normally part of the address, data, or command bus, or a conductor that normally carries the input load signal. The delay value might also be encoded on one or more conductors such as the conductors that normally form part of the address, data, or command bus.
A calibration process, which utilizes the feedback of delay element 90a, is preferably performed at regular intervals, to account for variations in temperature and voltage of the memory device. When timing events are supplied by an external source, it is desirable to provide such events on a periodically repeating basis for periodic recalibration. As will become more apparent in the more detailed discussion below, it is desirable to provide periodic bursts of such timing events for recalibration purposes. Such bursts can be provided concurrently with memory access cycles. However, it is not necessary for the timing signals to accompany individual memory cycles. Furthermore, the timing signals can be asynchronous to other signals used within the memory device.
In certain situations, it may be desirable to be able to configure, after device manufacture, the timing within a pipelined device such as described above to vary the time at which data is loaded within various pipeline stages. In the embodiment of
A delay block such as the one shown in
Second input load signal LD2, also referred to herein as a data register load signal, is used in place of the delay element outputs. The memory controller, which generates LD2, has its own timing elements that delay LD2 relative to LD1. During a read cycle, LD2 is gated to form a LOAD/ENABLE signal that loads read data register 62. Output driver 63 is responsive to this LOAD/ENABLE signal to present read data on the DATA bus. During a write cycle, LD2 is gated to form a LOAD/ENABLE signal that loads write data register 64 and enables driver 65. Timing details are similar to what is shown in
Address Interfacing Logic
On the left side of the
The LD, OP, and ADDR signals are received by input receiver blocks and buffers, labeled RB and RD. These blocks provide signal buffering and also impose uniform calibrated delays on the signals to ensure that the signals maintain their initial time relationships to each other. There are two versions: RB and RD. The RB is used for signals which need to drive relatively large loads. The RD is used for signals which need to drive relatively small loads. The specific design of the RB and RD blocks is discussed with reference to
The command bus supports the following operations in this simplified example:
OP[3]
OP[2]
OP[1]
OP[0]
Command
0
0
0
0
reserved
0
0
0
1
activate (ACT)
0
0
1
0
reserved
0
0
1
1
reserved
0
1
0
0
read (RD)
0
1
0
1
write (WR)
0
1
1
0
read and automatic pre-charge
(RDA)
0
1
1
1
write and automatic pre-charge
(WRA)
1
x
x
x
no operation
An actual DRAM product would include a richer set of operations. However, the set shown above is sufficient to demonstrate an implementation of the basic transactions for the asynchronous DRAM interface.
The ACT command accesses a row in the DRAM core 220, sensing it and storing it in the sense amplifier latches. The RD and RDA commands read a column of information (32 bits in this example design) from the sensed row. The WR and WRA commands write a column of information (32 bits) into the sensed row. The RDA and WRA commands also cause the accessed row to be pre-charged at the end of the column operation.
ADDR receives bank and row addresses for the ACT command, or bank and column addresses for the read and write commands. This design example includes two bits (ADDR[13:12]) for selecting one of four independent memory core banks. The ADDR[11:0] bits contain the row address or the column address, depending on the type of memory operation.
This example includes a sequence of address registers that receive addresses and that advance the addresses through the address registers in response to generated sequences of asynchronously timed register load signals. The memory core is responsive to the addresses after they have advanced through this sequence of address registers.
A first stage of address pipelining is supported by address register 226, which loads ADDR on the rising edge of LD1. Subsequent pipeline registers 227 and 228 receive successively delayed versions of column addresses (CLD1 and CLD2), and a final stage is supported in memory core 220 by an address register 230, which is loaded by another delayed version of the LD1 signal (COLLAT).
Row addresses and column addresses are handled differently. A row address is received along with the LD signal, and initially loaded by the LD1 signal in the first address pipeline register 226. A SENSE signal is generated by delay elements in response to the LD1 signal, at a predetermined time following the LD1 signal (see detailed timing below). The row address is received from first stage address register 226 by a memory core register 232, where the row address is loaded by the SENSE signal. Bank addresses for an automatic precharge operation (from a RDA or WRA command) are received from third stage address register 228 and are loaded into a bank address register 234 by yet another delayed version of LD1 (PLD). From there, the precharge operation's bank address is loaded by another delayed version of LD1 (PRECH) into the core's bank select register 236.
The described pipelined memory device includes memory timing or access logic 202 that is responsive to the input load signal LD1 and to the supplied 4-bit operation code to generate the delayed versions of LD1 mentioned above, and to thereby control the flow of information through the various pipeline registers shown in
Specifically, timing logic 202 generates the following signals:
The remaining signals, COLCYC, WR, WLD, QEN, QLD, and QMX are used in the data interface portion of the memory device, and will be described below with reference to
Registers 224, 226, 227, 228, and 234 each include a buffer for driving the load presented by the internal logic and wiring. The delay of these registers and buffers are masked by longer delays of control signals, so no timing calibration logic is used here.
The following table sets for exemplary timing parameters for the device of
Parameter
Value
Description
tRC
60
ns
Minimum time for successive activate
operations to the same bank
tRR
20
ns
Minimum time for successive activate
operations to different banks
tRP
15
ns
Minimum time between activate and pre-
charge operations to the same bank
tCC
10
ns
Minimum time for successive column
operations to a bank
tOP
5
ns
Minimum time for successive commands
on the OP bus
tBIT
2.5
ns
Minimum time to transport a bit on the
DQ, DM, DQS pins
tQ
2
ns
Maximum time from load signal to output
data valid for a register
tS
1
ns
Minimum time for a register input to be
valid prior to a load signal for a register
tH
1
ns
Minimum time for a register input to be
valid after a load signal for a register
tWROFF
5
ns
Maximum interval (either direction)
between the rising edges of LD and DQS
for WR
tCSH
20
ns
Minimum interval between rising edges of
SENSE and COLLAT
tCLS
5
ns
Minimum interval between rising edges of
COLLAT and COLCYC
tRCD
25
ns
Minimum interval between rising edges of
SENSE and COLCYC
tDAC
7.5
ns
Maximum interval from rising edge of
COLCYC to valid read data
tDOH
2.5
ns
Minimum interval from rising edge of
COLCYC to valid read data
tCPS
20
ns
Minimum interval between falling edge of
COLCYC and rising edge of PRECH
td
0.25
ns
Nominal delay of inverter pair in “d” cell
(adjustable)
tD
1.25
ns
Nominal delay of four inverter pairs and
buffer in “D” cell (adjustable)
tX
1.25
ns * X
Nominal delay of “NX” copies of the “D”
cell: tX = NX * tD
tREF and
10
ns
Nominal delay of “NREF” copies of the
tREF1
“D” cell: tREF = NREF * tD and NREF = 8
tTOT
2.50
ns
Nominal delay of “NTOT” copies of the
“D” cell: tTOT = NTOT * tD and NTOT = 2
tDEC
1.25
ns
Nominal delay of “NDEC” copies of the
“D” cell: tDEC = NDEC * tD and NDEC = 1
tLIM
1.25
ns
Nominal delay of “NLIM” copies of the
“D” cell: tLIM = NLIM * tD and NLIM = 1
tEXP
5.00
ns
Nominal delay of “NEXP” copies of the
“D” cell: tEXP = NEXP * tD and NEXP = 4
tEXP2
6.25
ns
Nominal delay of “NEXP2” copies of the
“D” cell: tEXP2 = NEXP2 * tD and NEXP2 = 5
tREP2
5.00
ns
Nominal delay of “NREP2” copies of the
“D” cell: tREP2 = NREP2 * tD and NREP2 = 4
tREP4
2.50
ns
Nominal delay of “NREP4” copies of the
“D” cell: tREP4 = NREP4 * tD and NREP4 = 2
A synchronous controller will probably generate the LD1 signal (and the other input signals). As a result, they will probably have an underlying regularity, and this is shown in
A first LD1 edge 270, in conjunction with an ACT command on the OP1 bus, initiates an activate operation. Along with the ACT command, a bank and row address Rx is presented on the ADDR1 bus. The rising edge 270 of LD1 loads the bank and row address into first-stage address register 226 and loads the ACT command into command register 224. The LD1 edge is delayed a time 2*tTOT, and then causes an edge on the SENSE signal. This loads the Rx address into core register 232 and starts the activate operation. No further activity occurs in the interface for this operation.
In this figure and subsequent figures, delays caused by delay elements are indicated by dashed lines from the event initiating the delay to the event resulting from the delay. In
A second LD1 edge 272 (received from the memory controller), in conjunction with an OP (RD or WR) command on the OP1 bus, initiates a column access operation. It is presented at a time top after the first LD1 edge. Along with the OP command, a bank and column address Cxa is presented on the ADDR1 bus. The second LD1 edge is delayed a time 2*tTOT, and then causes an edge on the CLD1 signal. This loads the Cxa address into second-stage pipeline register 227. The CLD1 edge is delayed an additional time t8, and then causes an edge on the CLD2 signal. This moves the Cxa from the pipeline register 227 into the third-stage pipeline register 228. The CLD2 edge is delayed an additional time t4, and then causes an edge on the COLLAT signal. This moves the Cxa from pipeline register 228 into the fourth-stage register 230 in the DRAM core. The COLLAT edge is delayed an additional time t4, and then causes an edge on the COLCYC signal. This signal controls data transport to and from the DRAM core, and will be discussed further with reference to
A third LD1 edge 277, in conjunction with an OP (RD or WR) command on the OP1 bus, initiates a second column access operation. This leading edge is presented a time tCC after the second LD1 edge 272. Again, a bank and column address Cxb is presented on the ADDR1 bus. The third LD1 edge 277 is delayed a time 2*tTOT, and then causes an edge on the CLD1 signal. This loads the Cxa address into second-stage pipeline register 227. The CLD1 edge is delayed an additional time t8, and then causes an edge on the CLD2 signal. This moves the Cxa from the pipeline register 227 into the third-stage pipeline register 228. The CLD2 edge is delayed an additional time t4, and then causes an edge on the COLLAT signal. This moves the Cxa from pipeline register 228 into the fourth-stage register 230 in the DRAM core. The COLLAT edge is delayed an additional time t4, and then causes an edge on the COLCYC signal. This signal controls data transport to and from the DRAM core, and will be discussed further with reference to
Note that other transactions could be presented to the DRAM while this first transaction is being processed. On the fourth and fifth LD1 edges 282 and 283, for example, ACT commands are directed to other banks in the DRAM. In this embodiment, these commands must be given a time tRR or more after the first ACT command. An ACT command directed to the first bank must be given a time tRC or more after the first ACT command.
Note also that there are several timing constraints imposed upon the timing of the COLLAT and COLCYC signals by the DRAM core. In particular, they must be issued a time tCSH and a time tRCD, respectively, after the SENSE signal.
Data Interfacing Logic
The data interfacing logic includes a write demultiplexer 240 (also referred to herein as demultiplexing logic) and a read multiplexer 242 (also referred to herein as multiplexing logic).
The write demultiplexer 240 accepts a sequence of four eight-bit words from DQ[7:0] and assembles them into a single 32-bit word (WD1 and WD) for writing to memory core 220. The assembled 32-bit word WD1 is loaded into an intermediate pipeline register 244, and then loaded into the appropriate memory core register 246 a subsequent, independent memory operation (see
The read demultiplexer 242 reads a 32-bit word RD[3:0][7:0] from the DRAM core read register 247 and splits it into four sequential eight-bit words for output from the memory device on DQ[7:0].
On the left side of
The right side of
The Write Demux block 240 accepts the write data DQ[7:0], the write mask DM, and the write data strobe DQS from the external memory controller component The DQS signal functions as a timing signal to load serially-received bits from DQ[7:0]. The signals are received by the RB and RD receiver cells 250 and 251. There are two versions: RB and RD. The RB cell is used for signals which need to drive relatively large loads. The RD cell is used for signals which need to drive relatively small loads. Both blocks have the same delay, controlled by calibration logic. These blocks are described with reference to
The DQS1 signal from the RB cell is used to clock a set of registers 254 which accumulate the four bits that appear serially on each wire for each write operation. One of these registers is loaded on the rising edge of DQS1, and the rest are loaded on the falling edge of DQS1. Toggle flip-flop 249 alternates its state between low and high on each falling edge of DQS2. It is forced to a low state by the RESET signal which is applied when the component is first powered on. The DQS2 signal is a delayed version of the DQS1 data strobe, using the delay element 245.
The result is that the four nine-bit serial words DQ3, DQ2, DQ1, and DQ0 will all be valid for a timing window surrounding the falling edge of DQS2 when the LD2 signal from toggle flip-flop 249 is high. These four serial words are loaded into register 241 on that falling DQS2 edge.
When the complete 36 bit parallel word (WD1[3:0][7:0] and WE[3:0]) is loaded into register 241, it is then driven and loaded into a intermediate pipeline register 244 on the rising edge of the WLD signal. The output of this register drives the WD[3:0][7:0] write data bus of the DRAM core. The DM bits are assembled on the WE[3:0] write mask bus in an identical manner.
The Read Mux block 242 accepts the read data RD[3:0][7:0] driven from the DRAM core after the rising edge of COLCYC. The parallel word is loaded into four eight bit registers 255 on the first rising edge of the QLD signal (when QMX is asserted to one). The four eight bit pieces are then shifted out serially onto the DQ[7:0] bus (when QMX is asserted to zero). The QEN signal is asserted to one enabling the output driver 258. Two-to-one multiplexers 256 are responsive to the QMX signal to control whether registers 255 are loaded from the RD[3:0][7:0] in response to the QLD signal, or are loaded from the previous register 255. Note that the pattern “1010” is appended to the RD[3:0][7:0] bus to form the timing signal on the DQS output. This timing information is treated like another data bit; the timing signals QLD and QMX shift the “1010” timing information onto the conductor used for the DQS signal.
This read data Qa is sampled by registers 255 at a time 2*tTOT+t24 after the second rising edge of LD1 (in the center of the valid window) by the first rising edge of the QLD control signal. The QMX and QEN control signals are asserted high a time 2*tTOT+t23 after the second rising edge of LD1. The QEN signal will remain asserted high for the time during which read data is being driven on the DQ and DQS pins. The QMX signal will remain high for the first rising edge of QLD, allowing the 32 bits of read data Qa[3:0][7:0] to be loaded into the serial output registers 255. The first eight bits Qa[3][7:0] will also be driven onto the DQ[7:0] pins a time tQ after the first rising edge of QLD. QMX will be left low for the next three QLD rising edges, allowing the remaining 24 bits Qa[2:0][7:0] to be shifted out.
The third LD1 edge 277 (with the RDA command) initiates a second column read operation. This command produces a second series of operations identical to that of the first column read, culminating in driving the second read data Qb[3:0][7:0] onto the DQ[7:0] pins. Note that the assertion of the QEN signal from the first read command merges with the assertion from the second read command; the QEN signal never returns to a low value between the commands.
The RDA command performs one set of operations not performed by the RD command; automatic pre-charge. The third rising edge 277 of LD1 causes the PLD signal to be asserted high at a time 2*tTOT+t24 later. This signal loads the Cxb bank address into a register 234 (
On the first rising edge of DQS in
The WLD control signal is delayed by 2*tTOT+t11 after the second rising edge 272 of LD1 (with the WR command). The rising edge of WLD causes register 244 to sample the WD1 and WE1 buses. This sampling point is designed to be in the center of the valid window for the data on these buses so that the offset parameter +tWROFF to −tWROFF has as much margin as possible. It is possible to adjust the delay path for the WLD signal if the sampling point needs to be shifted because of routing differences in the control and data wires for the memory subsystem.
The data on the WD and WE inputs to the DRAM core are sampled by register 246 (
On the third rising edge of DQS in
The WDA command performs one set of operations not performed by the WD command: automatic pre-charge. The third rising edge 277 of LD1 cause the PLD signal (
In the described embodiment, timing information is carried on a single, dedicated conductor corresponding to the DQS signal. However, in alternative embodiments such timing information might be encoded with the data itself. In such alternative embodiments, both timing information and data information might be transferred on a single signal line. A transmitter would receive a timing signal and the data signal, and in response produce a single signal to be carried by a single signal line to a receiver. In response, the receiver would separate the data information and timing information into two signals. A disadvantage of this technique is that the signal line must use some of its signaling bandwidth for the timing information. However, the technique might be desirable in some embodiments because it minimizes any skew between the data and timing information (as there would be if two separate signal lines were used).
Delay Element Calibration Circuit
The NodeA signal and MS bit of NodeB signal are passed through identical buffers 308 to give signals NodeC and Node E, respectively. NodeE is fed through a delay block 310, consisting of NREF copies of a D block. A D block is a delay element having a delay equal to tD, and will be described in more detail with reference to
CC block 312 compares the two signals on its INC and IND inputs and adjusts a pair of output voltages VREFP and VREFN so that the edges of the two signals are aligned in time. When a steady state voltage is reached, the delay between the pulses tREF1 of the TREF1 signal will match the delay tREF of the delay block NREF*D (to within the resolution supported by the CC block). The reference voltages can now be used to create calibrated delays within the interface logic.
Pulses are applied periodically on the TREF1 input from an external source such as the memory controller. Because of this, the reference voltages are periodically adjusted to compensate for process, voltage, and temperature variations. In this manner, an external delay reference can be used to create precise internal delays.
Note that it is not necessary that the TREF1 provide a continuous stream of pulses. Rather, short bursts of pulses are provided at regular intervals. The length of the interval is a function of how quickly temperature and supply voltage can change—this will typically be on the order of milliseconds. The length of the burst of pulses that are supplied will typically be on the order of 30 to 50 pulses—the CC block 312 in
Note that a simplified GEN block would consist of only the CC block and the delay block NREF*D. The TREF1 signal would be received by the RD block, and would drive the INC input and the input of the delay block. The TREF8 signal would simply be a buffered version of TREF1. The disadvantage of this simpler approach is its lack of robustness. The minimum and maximum delay range of tREF would be {0.5*tREF1, 1.5*tREF1}. If tREF ever acquired a value outside of this range (at power-up, for example), the CC block would drive the reference voltages in the wrong direction. The corresponding range of the more complicated GEN cell in
Compare and Control Block
There are four capacitors, which are charged to one of the two supply voltages when INC is high. They are CN+, CN−, CP+, and CP−. The capacitors each have a capacitance of “C”. When INC is low, two of the four capacitors dump their charge into the capacitors CREFP and CREFN on the reference voltage nodes VREFP and VREFN. These two capacitors have the capacitive values Nstep*C and Nstep*C. Thus, every time there is a pulse on INC, the reference voltages will make a step of (VDDA−GNDA)/Nstep in one direction or the other. At the steady-state reference voltages, the steps will alternate between up and down. The value of Nstep will be chosen as a compromise between the resolution of the steady state reference voltages and the time required to reach the steady state values at power-up time.
It would be possible to add logic to the CC block so that it would detect when it has made a series of steps in the same direction. It would then use a bigger capacitor to take bigger steps to the steady-state reference voltages. Once it began taking steps in the opposite direction, it would use the smaller capacitors for better delay resolution.
Note that VREFP and VREFN will always step in opposite directions. This will be clear when the details of the delay element are described (
Note also that the supply voltages VDDA and GNDA used by the CC block will be dedicated supplies that are different from the supplies used by the DRAM core and the data path logic of the interface. These dedicated supplies will be used only for the blocks of logic that generate precisely timed control signals. There will be less disturbance on these supplies due to switching noise, and the calibrated timing intervals will be more accurate as a result. The VDDA and GNDA will connect to the same external power supplies as the VDD and GND used by the rest of the DRAM, but will have dedicated pins and a dedicated set of power supply wires inside the component.
In both diagrams, the time when INC is high (after the INC rising edge), the four capacitors CN+, CN−, CP+, and CP− are charged to the supply rails. While this is happening, the output of the sampling register is settling to the value that determines what happens when INC drops low.
In the top diagram, the IND input doesn't have enough delay, and the INC rising edge samples IND as a “1”. This means that after INC drops low, the charge will be dumped so that VREFP is increased and VREFN is decreased.
In the bottom diagram, the IND input has too much delay, and the INC rising edge samples IND as a “0”. This means that after INC drops low, the charge will be dumped so that VREFP is decreased and VREFN is increased.
Note that the time that INC remains high and low doesn't affect the amount of charge dumped into the capacitors CREFP and CREFN on the reference voltage nodes VREFP and VREFN. It is only necessary to provide pulses on TREF1 with rising edges separated by the tREF1 interval—the duty cycle of these pulses is not critical.
Delay Elements
Each D cell 340 contains a plurality of “d” cell delay elements 342. Each d cell 342 is a pair of inverters 343 connected to VDDA through P transistors 344 whose gate voltage is VREFP, and connected to GNDA through N transistors 345 whose gate voltage is VREFN.
When VREFP increases, the resistance of the P transistors 344 increase, increasing the delay of a signal through the inverters 343. When VREFP decreases, the resistance of the P transistors 344 decreases, decreasing the delay of a signal through the inverters 343.
The behavior is complementary for an N transistor. When VREFN decreases, the resistance of the N transistors 345 increases, increasing the delay of a signal through the inverters 343. When VREFN increases, the resistance of the N transistors 345 decreases, decreasing the delay of a signal through the inverter 343.
At power-on, the VREFP and VREFN voltages are driven to the values of GNDA and VDDA, respectively. This corresponds to the shortest possible delay in the delay element. The GEN block 222 will drive VREFP higher and VREFN lower until the steady state values are reached. Note that VREFP and VREFN will always step in the opposite direction.
Other voltage-controlled delay structures are possible. The one that is described gives a good delay range with fairly modest area requirements. It would also be possible to use a digitally-controlled delay structure, in which delay elements were added and removed with a multiplexer structure. This would yield much coarser delay resolution, however. A hybrid delay unit with a coarse structure and a fine structure could also be used.
D cell 340 also includes a buffer 350 (inverter pair) for restoring the nominal slew rate to a signal passing through the block. This permits the D cell to drive a larger load directly. The delay of the “D” cell is tD=n*td, where td is the “d” cell delay.
Receiver Blocks
Each receiver block has a real signal path, shown in the upper part of
The real signal path consists of an input signal IN passing through a receiver 360(a), a delay cell 362(a) comprising NADJ D cells, and a buffer 364(a) to the output OUT.
The image signal path consists the TREF8 signal (from the GEN block of
The reference voltages VADJP and VADJN produced by the CC block control the delay of the identical NADJ*D blocks 362(a) and 362(b). As a result, the pulses from the TREF8 signal will propagate through the two paths in the lower block, and will be compared in the CC cell 366. The CC cell will adjust the VADJP and VADJN voltages to make the delay of the receiver 360(b), delay cell 362(b), and buffer 364(b) equal to tTOT.
In the upper cell, the delay seen by the input signal IN through the receiver 360(a), delay cell 362(a), and buffer 364(a) will also be equal to tTOT since all the components are matched and the VADJP and VADJN voltages are shared. If the delay of the receiver and buffer change because of temperature and supply voltage variations, the delay of the NADJ*D delay cell will change in a complementary fashion so the overall delay remains tTOT.
This example implementation of an asynchronous DRAM interface assumes that the real signal path of each RB and RD cell has a dedicated image or reference signal path. In an actual implementation, it is likely that the image signal paths could be shared among all real signal paths that are matched. For example all the bits of the address input ADDR[13:0] could share one image path. This would reduce the cost of calibrating the RB and RD delay to the area of the (NADJ*D) delay cell plus a fraction of the image signal path cell. The VADJP and VADJN voltage signals would be routed to all the (matched) real signal paths from the image signal path.
It would also be possible to use the real signal path to generate its own adjustment voltage. This requires that the real signal path consist of pulses with a repetition rate that is constrained by the logic in the CC block. The advantage of this is that the delays are measured and adjusted in the real signal path, saving some area and perhaps making the timing calibration more accurate. The disadvantage is that if a real path is not exercised often enough, its delay may drift. The advantage of the image signal path is that it can have its adjustment voltage updated without interfering with its real signal operation.
Timing Logic
There are five DEC blocks 401 which decode the four bit command OP2 into five command signals, indicating an activate operation (ACT), a column operation (RD/WR/RDA/WRA), a column read operation (RD/RDA), an automatic pre-charge operation (RDA/WRA), and a column write operation (WR/WRA).
These five signals then pass through a number of delay cells 402, each of which has a delay that is indicated in the figure. For example, the cell “Nx*D” generates the delay tx=Nx*tD=X*tD, where the value of “X” can be {1, 4, 8, 11, 23, 24}. These delay cells use the standard reference voltages VREFP and VREFN because the delays are calibrated to the reference delay tD from the GEN cell. The EXP, REP2, and REP4 (each of which will be described below) then shape the decoded and delayed signals cells.
The LD1 load signal passes through a delay block 406(a) (NDEC*D). This provides a delay of tDEC=NDEC*tD which will be enough to match the load-to-output delay of the OP2 register 224 (
Below the real signal path is the image signal path. It consists of the TREF8 signal (from the GEN block of
The reference voltages VADJP and VADJN produced by the CC block 414 control the delay of the NADJ*D blocks. As a result, the pulses from the TREF8 signal will propagate through the two paths in the lower block, and will be compared in the CC cell. The CC cell will adjust the VADJP and VADJN voltages to make the delay of the two delay cells 406(b), 410(b) and buffers 412(b) equal to tTOT.
In the upper cell, the delay seen by the input signal IN through the delay cell 406(a), and gate 408(a), delay cell 410(a), and buffer 412(b) will also be equal to tTOT since all the components are matched and the VADJP and VADJN voltages are shared. If the delay of the receiver and buffer change because of temperature and supply voltage variations, the delay of the NADJ*D delay cell will change in a complementary fashion so the overall delay remains tTOT.
This example implementation of an asynchronous DRAM interface assumes that the real signal path of each DEC cell has a dedicated image signal path. In an actual implementation, it is likely that the image signal paths could be shared among all real signal paths that are matched. This is particularly easy since each DEC cell fans out to either one or two other cells that are also part of the Decode block. This would reduce the cost of calibrating the DEC delay to the area of the (NADJ*D) delay cell plus a fraction of the image signal path cell. The VADJP and VADJN voltage signals would be routed to all the (matched) real signal paths from the image signal path.
The NodeC signal is expanded to the appropriate width by the next circuit. NodeC passes to the “set” input of an SR latch 446(a). This causes the “q” output to be set high. NodeC also passes through a delay block 448 (NEXP*D) which provides a delay of tEXP=NEXP*tD. The delayed signal passes to the “reset” input of the SR latch 446(a), causing the “q” to return low after a pulse width of about tEXP.
The NodeF output of the SR latch 446(a) passes through a third delay block 450(a) (NADJ*D) and a buffer 452(a) which drives the control signal to the interface logic and the DRAM core. This third delay line is used to add an adjustable delay so the total delay of the EXP block remains fixed at the desired value tTOT.
Below the real signal path is an image signal path. It consists of the TREF8 signal (from the GEN block) passing through an identical “and” gate 440(b), SR latch 446(b), delay cell 450(b) (NADJ*D) and buffer 452(b). The buffer for the image signal drives a load that is equivalent to that driven by the buffer for the real signal. This image signal is fed into the IND input of a CC block 454. The TREF8 signal also goes through a second delay cell 456 with a delay of tTOT=NTOT*tD and is fed into the INC input of the CC block. The reference voltages VADJP and VADJN produced by the CC block 454 control the delay of the NADJ*D blocks 450(a) and 450(b).
The pulses from the TREF8 signal propagate through the two paths in the lower block, and are compared in the CC cell 454. The CC cell adjusts the VADJP and VADJN voltages to make the delay of the two delay cells and buffer equal to tTOT. Note that the delay cells (NLIM*D) and (NEXP*D) are not included here because there is no need to shape the TREF8 reference signal; the CC block only uses the relative positions of the IND and INC rising edges to generate the adjustment voltage.
In the upper cell, the delay seen by the input signal IN through the “and” gate, SR latch, delay cell and buffer will also be equal to tTOT since all the components are matched and the VADJP and VADJN voltages are shared. If the delay of the receiver and buffer change because of temperature and supply voltage variations, the delay of the NADJ*D delay cell will change in a complementary fashion so the overall delay remains tTOT.
Note also that the pulse width at NodeJ is (tEXP−tNOR). The pulse width will have some variation with respect to temperature and voltage since the tNOR delay is uncalibrated. However, the position of the falling edge of all control signals is not important—it is only necessary to precisely position the rising edges. Thus, this slight variation of pulse width will not affect the performance of the memory component.
This example implementation of an asynchronous DRAM interface assumes that the real signal path of each EXP cell has a dedicated image signal path. In an actual implementation, it is likely that the image signal paths could be shared among all real signal paths that are matched. This could be accomplished by adding dummy loading to the real signals so that all EXP blocks see the same effective load. This would reduce the cost of calibrating the DEC delay to the area of the (NADJ*D) delay cell plus a fraction of the image signal path cell. The VADJP and VADJN voltage signals would be routed to all the (matched) real signal paths from the image signal path.
The NodeC signal is expanded to the appropriate width by the next circuit. NodeC passes to the “set” input of an SR latch 464(a). This causes the “q” output to be set high. NodeC also passes through a delay block 466 (NEXP2*D) which provides a delay of tEXP2=NEXP2*tD. The delayed signal passes to the “reset” input of the SR latch, causing the “q” to return low after a pulse width of about tEXP.
The NodeF output of the SR latch 464(a) passes through an “or” gate 468(a). The NodeF signal also passes through a delay block 470 (NREP2*D). This provides a delay of tREP2=NREP2*tD. The delayed NodeF signal and the undelayed NodeF signal are or'ed to give NodeH. The values of tEXP2 and tREP2 are chosen so that the two pulses overlap and merge. This is because the REP2 block produces the enable signal for the output driver. It must remain asserted (without glitching low) during the whole time that read data is driven.
The NodeH output of the “or” gate passes through a third delay block 472(a) (NADJ*D) and a buffer 474(a) which drives the control signal to the interface logic and the DRAM core. This third delay line is used to add an adjustable delay so the total delay of the REP2 block remains fixed at the desired value tTOT.
Below the real signal path is an image signal path. It consists of the TREF8 signal (from the GEN block of
The pulses from the TREF8 signal will propagate through the two paths in the lower block, and will be compared in the CC cell 478. The CC cell will adjust the VADJP and VADJN voltages to make the delay of the two delay cells and buffer equal to tTOT. Note that the delay cells (NLIM*D), (NEXP2*D) and (NREP2*D) are not included here because there is no need to shape the TREF8 reference signal; the CC block only uses the relative positions of the IND and INC rising edges to generate the adjustment voltage.
In the upper cell, the delay seen by the input signal IN through the “and” gate, SR latch, delay cell, “or” gate, and buffer will also be equal to tTOT since all the components are matched and the VADJP and VADJN voltages are shared. If the delay of the receiver and buffer change because of temperature and supply voltage variations, the delay of the NADJ*D delay cell will change in a complementary fashion so the overall delay remains tTOT.
Note also that the pulse width at NodeJ is (tEXP2+tREP2−tNOR). The pulse width will have some variation with respect to temperature and voltage since the tNOR delay is uncalibrated. However, the position of the falling edge of all control signals is not important—it is only necessary to precisely position the rising edges. Thus, this slight variation of pulse width will not affect the performance of the memory component.
If the subsequent column operation is also a RD or RDA command, there will be another pulse on NodeA a time tCC after the first pulse (dotted line). The pulse that is produced a time tTOT later on NodeJ will be merged with the first pulse because of the “or” gate that drives NodeH. This ensures that the output driver remains on when driving read data from consecutive read accesses.
This example implementation of an asynchronous DRAM interface assumes that the real signal path of the REP2 cell has a dedicated image signal path (i.e., only one REP2 cell is used in this example). Other implementations might use more than one REP2 cell, in which case the image signal paths could be shared among all real signal paths that are matched. This could be accomplished by adding dummy loading to the real signals so that all REP2 blocks see the same effective load. This would reduce the cost of calibrating the DEC delay to the area of the (NADJ*D) delay cell plus a fraction of the image signal path cell. The VADJP and VADJN voltage signals would be routed to all the (matched) real signal paths from the image signal path.
The NodeC output of the pulse limiter passes through an “or” gate 504(a). The NodeF signal also passes through three delay blocks 506, 507, and 508 (NREP4*D). Each provides a delay of tRP4=NREP4*tD. The three delayed NodeF signals and the undelayed NodeF signal are or'ed at gate 504(a) to give NodeH. The values of tLIM and tREP4 are chosen so that the four pulses do not overlap. This is because the REP4 block produces the load signal for the output registers 255 (
The NodeH output of the “or” gate 504(a) passes through a third delay block 510(a) (NADJ*D) and a buffer 512(b) which drives the control signal to the interface logic and the DRAM core. This third delay line is used to add an adjustable delay so the total delay of the REP4 block remains fixed at the desired value tTOT.
Below the real signal path is the image signal path. It consists of the TREF8 signal (from the GEN block of
The pulses from the TREF8 signal will propagate through the two paths in the lower block, and will be compared in the CC cell. The CC cell will adjust the VADJP and VADJN voltages to make the delay of the two delay cells and buffer equal to tTOT. Note that the delay cells (NLIM*D) and (NREP4*D) are not included here because there is no need to shape the TREF8 reference signal; the CC block only uses the relative positions of the IND and INC rising edges to generate the adjustment voltage.
In the upper cell, the delay seen by the input signal IN through the “and” gate, “or” gate, delay cell, and buffer will also be equal to tTOT since all the components are matched and the VADJP and VADJN voltages are shared. If the delay of the receiver and buffer change because of temperature and supply voltage variations, the delay of the NADJ*D delay cell will change in a complementary fashion so the overall delay remains tTOT.
The initial pulse on NodeA becomes four pulses, the first delayed by tTOT, the rest following at intervals of tREP4. Each pulse is asserted for tLIM.
If a subsequent column operation is also a RD or RDA command, there will be another pulse on NodeA a time tCC after the first pulse (dotted line). The pulse that is produced a time tTOT later on NodeJ will be NodeA a time tCC after the first pulse. The minimum tCC value will be equal to 4*tREP4.
This example implementation of an asynchronous DRAM interface assumes that the real signal path of the REP4 cell has a dedicated image signal path (i.e., only one REP4 cell is used in this example). Other implementations might use more than one REP4 cell, in which case the image signal paths could be shared among all real signal paths that are matched. This could be accomplished by adding dummy loading to the real signals so that all REP4 blocks see the same effective load. This would reduce the cost of calibrating the DEC delay to the area of the (NADJ*D) delay cell plus a fraction of the image signal path cell. The VADJP and VADJN voltage signals would be routed to all the (matched) real signal paths from the image signal path.
As shown in
The multiplexer 602 has four inputs, which receive versions of the LD1 signal that have been delayed by successively larger intervals by delay elements 604. The value of Qsel[3:0] will enable an undelayed signal, or will enable one of three delayed versions of the signal, with incremental delays of 1*tD, 2*tD, and 3*tD. This will cause all four of the affected signals to shift together in time, causing the read data bit windows on the external pins of the DRAM device to shift.
Aspects of embodiments described herein are set forth for example and without limitation in the following numbered clauses:
1. A method of controlling a memory device, the method comprising:
14. A method of controlling a memory device, the method comprising:
Although details of specific implementations and embodiments are described above, such details are intended to satisfy statutory disclosure obligations rather than to limit the scope of the following claims. Thus, the invention as defined by the claims is not limited to the specific features described above. Rather, the invention is claimed in any of its forms or modifications that fall within the proper scope of the appended claims, appropriately interpreted in accordance with the doctrine of equivalents
Tsern, Ely K., Hampel, Craig E., Ware, Frederick A., Stark, Donald C.
Patent | Priority | Assignee | Title |
11482273, | Nov 11 2020 | XILINX, Inc. | Strobe tree circuit for capturing data using a memory-sourced strobe |
8059476, | Feb 28 2001 | Rambus Inc. | Control component for controlling a delay interval within a memory component |
8295107, | Feb 28 2001 | Rambus Inc. | Asynchronous pipelined memory access |
Patent | Priority | Assignee | Title |
5233557, | Jun 29 1990 | MATSUSHITA ELECTRIC INDUSTRIAL CO , LTD , | Memory control device |
5661688, | Apr 24 1995 | SAMSUNG ELECTRONICS CO , LTD | Semiconductor memory device with an extended data output mode |
5835443, | Sep 20 1995 | Renesas Electronics Corporation | High speed semiconductor memory with burst mode |
5848260, | Dec 10 1993 | DASSAULT SYSTEMES SIMULIA CORP, | Computer system for simulating physical processes |
5886948, | Dec 20 1996 | Round Rock Research, LLC | Memory having a plurality of external clock signal inputs |
5923615, | Apr 17 1998 | SHENZHEN XINGUODU TECHNOLOGY CO , LTD | Synchronous pipelined burst memory and method for operating same |
6016282, | May 28 1998 | Round Rock Research, LLC | Clock vernier adjustment |
6075730, | Oct 10 1997 | Intel Corporation | High performance cost optimized memory with delayed memory writes |
6101136, | Jan 13 1998 | Longitude Licensing Limited | Signal delay device for use in semiconductor storage device for improved burst mode operation |
6101138, | Jul 22 1999 | Eton Technology, Inc. | Area efficient global row redundancy scheme for DRAM |
6154821, | Mar 10 1998 | Rambus, Inc | Method and apparatus for initializing dynamic random access memory (DRAM) devices by levelizing a read domain |
6240045, | May 06 1999 | Renesas Electronics Corporation | Synchronous semiconductor integrated circuit capable of improving immunity from malfunctions |
6266285, | Apr 18 1990 | Rambus Inc. | Method of operating a memory device having write latency |
6298004, | Aug 13 1996 | SOCIONEXT INC | Semiconductor device, semiconductor system, and digital delay circuit |
6314051, | Apr 18 1990 | Rambus Inc. | Memory device having write latency |
6327196, | Feb 27 1998 | Round Rock Research, LLC | Synchronous memory device having an adjustable data clocking circuit |
6351433, | Oct 15 1999 | Renesas Electronics Corporation | Semiconductor memory device employing pipeline operation with reduced power consumption |
6401167, | Oct 10 1997 | Rambus Incorporated | High performance cost optimized memory |
6449727, | May 07 1998 | Kabushiki Kaisha Toshiba | High-speed data transfer synchronizing system and method |
6539454, | Apr 01 1998 | CONVERSANT INTELLECTUAL PROPERTY MANAGEMENT INC | Semiconductor memory asynchronous pipeline |
6640292, | Sep 10 1999 | Rambus Inc | System and method for controlling retire buffer operation in a memory system |
6643787, | Oct 19 1999 | Rambus Inc | Bus system optimization |
6842864, | Mar 10 1998 | Rambus Inc. | Method and apparatus for configuring access times of memory devices |
7362626, | Feb 28 2001 | Rambus Inc. | Asynchronous, high-bandwidth memory component using calibrated timing elements |
WO9950852, |
Executed on | Assignor | Assignee | Conveyance | Frame | Reel | Doc |
Mar 03 2008 | Rambus Inc. | (assignment on the face of the patent) | / |
Date | Maintenance Fee Events |
Nov 05 2012 | M1551: Payment of Maintenance Fee, 4th Year, Large Entity. |
Dec 16 2016 | REM: Maintenance Fee Reminder Mailed. |
May 05 2017 | EXP: Patent Expired for Failure to Pay Maintenance Fees. |
Date | Maintenance Schedule |
May 05 2012 | 4 years fee payment window open |
Nov 05 2012 | 6 months grace period start (w surcharge) |
May 05 2013 | patent expiry (for year 4) |
May 05 2015 | 2 years to revive unintentionally abandoned end. (for year 4) |
May 05 2016 | 8 years fee payment window open |
Nov 05 2016 | 6 months grace period start (w surcharge) |
May 05 2017 | patent expiry (for year 8) |
May 05 2019 | 2 years to revive unintentionally abandoned end. (for year 8) |
May 05 2020 | 12 years fee payment window open |
Nov 05 2020 | 6 months grace period start (w surcharge) |
May 05 2021 | patent expiry (for year 12) |
May 05 2023 | 2 years to revive unintentionally abandoned end. (for year 12) |